1

Edge Intelligence for Empowering IoT-basedHealthcare Systems

Vahideh Hayyolalam, Member, IEEE, Moayad Aloqaily, Member, IEEE, Oznur Ozkasap, SeniorMember, IEEE, Mohsen Guizani, Fellow, IEEE

F

Abstract—The demand for real-time, affordable, and efficient smarthealthcare services is increasing exponentially due to the technologicalrevolution and burst of population. To meet the increasing demands onthis critical infrastructure, there is a need for intelligent methods to copewith the existing obstacles in this area. In this regard, edge computingtechnology can reduce latency and energy consumption by movingprocesses closer to the data sources in comparison to the traditionalcentralized cloud and IoT-based healthcare systems. In addition, bybringing automated insights into the smart healthcare systems, artificialintelligence (AI) provides the possibility of detecting and predicting high-risk diseases in advance, decreasing medical costs for patients, andoffering efficient treatments. The objective of this article is to highlightthe benefits of the adoption of edge intelligent technology, along withAI in smart healthcare systems. Moreover, a novel smart healthcaremodel is proposed to boost the utilization of AI and edge technology insmart healthcare systems. Additionally, the paper discusses issues andresearch directions arising when integrating these different technologiestogether.

Index Terms—Smart healthcare, Edge computing, IoT, Artificial intelli-gence, Machine learning.

1 INTRODUCTION

Healthcare is one of the crucial parts of human life. Dueto the rapid population growth and the rise of various dis-eases, there exists a drastic surge in requests for healthcareand its facilities. Also, with the technological evolution andthe emergence of Internet of things (IoT) coupled with theimprovement of next-generation wireless communications,the concept of smart healthcare or connected healthcaresystems has been introduced as a developed version ofconventional healthcare systems. In fact, smart healthcarerefers to a health service that exploits technologies suchas IoT, wearable devices, and advanced communicationprotocols to dynamically connect patients, caregivers, andhealth institutions and transfer information among them [1].Smart healthcare services provide the possibility of handlingand responding to medical requests remotely in an intel-ligent way, which diminishes hospitalization remarkablyand helps people and doctors to predict, detect, diagnose,and treat diseases intelligently. Also, smart healthcare is

• V. Hayyolalam and O. Ozkasap are with Koc University, Istanbul, Turkey.E-mails: {vhayyolalam20; oozkasap}@ku.edu.tr.

• M. Aloqaily is with xAnalytics Inc., Canada. E-mail: malo-qaily@xanalytics.ca.

• M. Guizani is with Qatar University, Qatar. E-mail: mguizani@ieee.org.

used for preventing and controlling the outbreak of con-tagious and infectious diseases such as the Ebola virus,Avian influenza, Chickungunya virus [2], and recently theCovid-19 pandemic. Taking all the above mentioned issuesinto account, adoption of efficient smart healthcare serviceswould certainly improve the health level of our society.

In order to expand the healthcare facilities and satisfythe vast amount of users’ requests, smart healthcare sys-tems employ numerous smart devices and IoT technolo-gies. Thus, a massive amount of heterogeneous medicalrecords generated by smart devices need to be processedand evaluated accurately based on the target of the system[3]. Therefore, there is a need for effective communicationsamong various entities of smart healthcare community anddata centers to ensure low response time for emergencycases in real-time health needs, since the high response timeand high latency at data centers are critical risks in smarthealthcare systems and can lead to the irreparable catas-trophes [1]. Keeping these facts in mind, edge computingtechnology, and AI are the best choices for these issues. Es-pecially, adopting the combination of these two technologiescan pave the way for solving various challenges in the scopeof smart healthcare.

Edge computing technology moves some processes closeto the data sources, which reduces the load of transmitteddata conspicuously [4]. In fact, edge technology enablessmart healthcare systems to conduct some processes andstore some data near the end-users rather than transferringall the records to the cloud data centers [5]. In this way,smart healthcare systems only need to send the results ofprocesses and some raw data to the remote cloud data cen-ters. Various research works by utilizing edge technology inthe field of smart healthcare can conduct big data analysis,store and process sensitive medical data, diminish latency,reduce energy consumption, reduce cost, reduce networkcongestion, and response time.

Moreover, by emulating human cognition in analyzingdata through complex algorithms, AI can smooth the pathof estimating the conclusions without the direct involve-ment of human. In fact, AI is responsible for investigatingrelations between the treatment, prevention, or detectiontechniques of diseases based on the collected data frompatients In the edge context, the resource constraints ne-cessitate innovative and lightweight methods to be able toexecute them at the edge environment. Therefore, Machine

arX

iv:2

103.

1214

4v1

[cs

.LG

] 2

2 M

ar 2

021

2

Learning (ML) and Deep Learning (DL) as the subsets ofAI are mostly preferred in edge technology to train thesystem for learning, making decisions, and actuating. Theintegration of edge computing and AI introduces an edgeintelligence concept, which leads to emerge a revolutionin smart healthcare systems [6]. Researchers have proposedseveral smart healthcare systems adopting AI methods suchas DL, Deep Neural Network (DNN), reinforcement learn-ing (RL), K-means , and other ML-based algorithms.

The contributions of this article can be summarized asfollows:

• Classifies the current AI-assisted edge-based smarthealthcare solutions via three viewpoints, namely,the application cases in IoT smart healthcare, theproper AI deployment location at the edge, and themost suitable type of adopted AI technique.

• Proposes a novel smart healthcare model leveragingAI and edge technology together to reduce the limi-tations of the existing models.

• Adopts the use of deep reinforcement learning(DRL). According to the number of vital sign typesneeded to be monitored and the amount of collectedmedical data, the proposed system is designed todistribute the DRL layers and processes optimally.

• Generalize the design of the proposed DRL modelin a way it can be used in parallel mode at theedge layer. It means that the processes with the samefunctionality are distributed among different devicesor DRL layers to reduce response time, latency, andcongestion.

The magazine is organized as follows, Section 2 dis-cusses how edge computing affects smart healthcare sys-tems. Then, Section 3 states the AI implication in edge-basedsmart healthcare systems as well as the pros and challengesof integrating AI and edge technology. Section 4 introducesa new framework for smart healthcare exploiting AI andedge technology. Open issues and challenges are discussedin Section 5. Eventually, Section 6 provides the conclusionof this research.

2 EMERGING EDGE TECHNOLOGY IN SMARTHEALTHCARE

Many healthcare systems have adopted cloud comput-ing to deliver affordable and scalable solutions in orderto process and store massive amounts of data recordedvia numerous biosensors. However, since cloud-orientedhealthcare systems mostly include mobile devices, network,and cloud servers, there are often long distances betweenthe systems’ units, which intensifies a high latency issue inthese systems [5]. Thus, emergency and real-time medicaldemands cannot rely on cloud-based healthcare systems.Furthermore, the vast amounts of records produced bysensors need to regularly be transferred to the cloud forbeing processed and stored, which leads to high energyconsumption and high cost. Also, most of chronic patientsrequire low-cost mobile environments, which is not sup-ported by cloud-oriented solutions [3].

Edge computing as an additional layer between end-users’ devices and remote cloud data centers is an emergingsolution to prevail cloud-based healthcare limitations. In

fact, edge computing is a smart gateway that plays therole of a bridge between devices and cloud data servers [7].Edge-assisted solutions transfer processes close to the end-users, which provides short response time and reduces en-ergy usage. Indeed, edge technology enables smart health-care systems to mine and process the collected data on theedge servers and devices near the user rather than trans-ferring it to the cloud servers. Processing at the edge alsoassists and boosts security, privacy, mobility, low networkbandwidth usage, geographical distribution, and locationawareness and facilitates online diagnosis and analytics,which reduces clinic and hospital visits [3]. Hence, edgetechnology contributes to the development of smart health-care services by providing swift, more comprehensive, andubiquitous treatments.

Sen

sors

Ed

ge

dev

ices

Ed

ge

serv

ers

Clo

ud

da

ta c

ente

r

Sensor types

Implanted

Wearable

Collected data types

Heart beat rate

Body temperature

Electroencephalogram

Respiration rate

Blood Glucose

Electrocardiogram

Electromyogram

Blood pressure

Data(vital sign) collection Data cleaning

Functionalities

Functionalities

Powerful data processing Permanent Data storage

Functionalities

Data pre-processing

Temporary data storage

Data aquisition

Data integration

Data transformation

Data reduction

Data generalization

Sleep quality

Fig. 1: Edge-assisted smart healthcare functionality flow.

By and large, edge computing architecture in smarthealthcare applications consists of four layers, includingsensors layer, edge devices layer, edge servers’ layer, andcloud data center layer. Figure 1 illustrates this architectureas a functionality flow, which indicates the functionality andthe common devices used in each layer. The sensor layercontains different sensors, including implanted and wear-able sensors that are responsible for collecting vital signsand data from patients. We indicate some data types thatcan be collected via sensors in a smart healthcare system.The sensor layer communicates with the edge devices layervia short range and low power wireless communicationprotocols such as ZigBee, Bluetooth, and Wi-Fi [5]. Edgedevices layer includes users’ devices such as smartphones,raspberry pi, tablets, and smart watches, which are capable

3

of conducting some processes and short-term storing data.The information is sent via Bluetooth/ Wi-Fi/internet fromedge devices layer to edge servers’ layer. Edge servers’ layercomprises micro data centers and is able to conduct moreprocesses and stores more data in comparison to the edgedevices layer [1]. As shown in Figure 1, the functionality ofedge devices layer and edge servers layer are similar. Theonly difference is that the storage capacity and the compu-tational power in edge server layer is more powerful thanedge device layer. Cloud data centers’ layer encompassesmore powerful processors and has the ability to store thevast amount of data. This layer communicates with edgeservers’ layer via the Internet.

3 AI IN EDGE ASSISTED SMART HEALTHCARESYSTEMS

This section first provides an overview and benefitsof applying AI in edge assisted smart healthcare systems.Then, some related research works are examined in detail.Three different classifications of the investigated works areprovided and illustrated in Figure 2 and Figure 3. Finally, abrief discussion along with Table 1 including the summaryof inspected papers, are brought in this section.

The first classification considers the AI deployment placein edge computing to classify the papers. Regarding this,papers are classified into two groups, including edge devicedeployment and edge server deployment. The second classifica-tion categorizes the research works based on their appliedcases in smart healthcare into five groups (see column appli-cation case in Table 1). The mentioned five groups include1. Neurological disorders (e.g., Epileptic seizure, chronic braindisorders), 2. Activity monitoring (e.g., Elderly monitoring,fitness tracking, emergency care, and tele-health advice), 3.Behavior recognition (e.g., stress detection, feeling recogni-tion), 4. Cardiovascular disorders (e.g., Heart problems, bloodpressure), 5. Infection outbreak (e.g., various viruses such asChickungunya, COVID 19). These two classifications areoutlined through Figure 2, which illustrates the adoptedalgorithms, their associated healthcare cases, and their de-ployment places on edge. For instance, the DNN algorithmis employed for behavior recognition, which is deployedat an edge device. However, it can be seen that three ofpapers (CNN, SVM[7], K-means [8], and HTM [9]) are notassigned to any of the application case groups, since theirapplication healthcare cases have not been clarified by theauthors. Moreover, Figure 3 depicts the third categorization,which is based on the adopted AI technique, including threesubcategories: ML, DL, and hybrid methods. The ML-basedtechniques, including supervised learning, unsupervised learn-ing, and RL, and also, hybrid technologies encompassingthe hybrid of ML and DL, the hybrid of ML and heuristicmethods.

3.1 Overview and Benefits

The integration of AI and edge technology enables smarthealthcare systems to run AI techniques locally on end-user devices where data are produced, while in traditionalsystems data processing procedures have been performedby AI techniques placed in the cloud far from end-users.This integration provides the possibility of processing dataon edge devices in a few milliseconds, which delivers real-

Fig. 2: The classification of state-of-the-art solutions basedon the AI deployment place and their application case.

time results. Since healthcare services are directly related tohuman health, supplying real-time services and results is socrucial in critical conditions [10].

Researchers have recently focused on moving the infer-ence portion of the AI workflow to the edge devices. Thesmart healthcare system can conduct the training portion inone environment and run the other parts in another place. Inorder to train ML/DL, the vast amount of data and powerfulcomputing items are needed; thus, the cloud is more suitablefor this part of AI techniques. Whereas, the inference part orrunning the trained models in novel records can be executedat edge devices. Moreover, some techniques such as modelcompression methods by reducing large AI models intoslight ones can move some training to the edge devicesgradually [5].

All in all, the adoption of AI techniques for processingmedical data on the edge of the network can contributeto providing smarter healthcare services with prompt re-sponse, high privacy, high accuracy, improved reliability,enhanced bandwidth usage, low power consumption andlow latency. Also, it can reduce the burden on the networkby diminishing the amount of transmitted data to the cloudservers [5].

3.2 Review and Classification

In this subsection, AI-based edge assisted smart healthcaresolutions are discussed in two groups according to the firstclassification mentioned earlier (See figure 2).

3.2.1 AI in edge devices

Researchers in [6] have outlined a novel architecture forhuman activity recognition that utilizes Dockers technologyon edge devices accompanied by AI techniques. The pro-posed architecture consists of sensors network, central cloudserver controlling the sensors, and edge computing as abridge between sensors and cloud. The regularized extreme

4

learning machine (RELM) is used in edge computing toclassify the collected data from the sensors.

Also, in [1], the authors have proposed a context-awareedge-assisted smart healthcare framework by exploiting AItechniques. They have adopted DL for data compressionin mobile edge devices to diminish bandwidth and energyusage. They have employed stacked auto-encoders (SAE),which is a type of neural network and hierarchically extractsdata. Also, in order to guarantee high reliability and a swiftresponding in case of detecting or predicting disorders inepileptic seizure detection, the authors have adopted a dataclassification method leveraging ML algorithms.

Researchers in [11] have proposed a human activityprediction system adopting edge technology and AI tech-niques. In the proposed system, data is collected by sensorsand transferred to the edge device (laptop), where featuresare extracted from the collected data. Then, with regards tothe features, a Recurrent Neural Network (RNN) is trainedon the edge device. The trained RNN is utilized for predict-ing human activities.

AI

in

ht cl aa re eH Smart Systems

MachineLearning

DeepLearning

Hybrid

ML+DLML+

Heuristics

ReinforcementLearning

SupervisedLearning

UnsupervisedLearning

RL

RELM

SAE+ML

SVM

FKNNNN

RNN

DNN

CNN+SVM

HTM

K-means

NB+WOA

LRA+DTC+FOA

Fig. 3: The classification of state-of-the-art solutions basedon the adopted AI techniques.

Authors in [12] have proposed a smart healthcare sys-tem using edge technology and DL for real-time stresslevel detection by taking physical activity into account. Theproposed system adopts wrist band sensors to collect therequired data. Also, DNN is utilized on edge devices. Inthe employed DNN, the input layer receives the collecteddata from sensors. Some hidden layers are responsible fordata and stress level analyzing, the output layer reportsthe detected stress level, and eventually, the results arestored in the cloud. In the proposed network, a supervisedlearning method is adopted for data training, the gradientdescent algorithm is employed for optimizing the trainingalgorithm, and logistic regression is used for classifying ofstress levels.

Also, researchers in [13] have designed an offloadingmodel for edge-aided smart healthcare systems adopting

AI methods to preserve user privacy in emergency careand Telehealth advice application. The proposed model usesthe RL-based offloading algorithm to protect the privacyof healthcare data during the processing procedure. Re-searchers in [14] have designed a real-time elderly monitor-ing system on a heterogeneous multi-core edge device usingAI techniques. The proposed system utilizes the shimmerdevice for data acquisition; then, the collected data is com-pressed by a compressive sensing (CS) method. Next, thecompressed data is transferred to the edge gateway throughBluetooth. The ODROID XU4 is used as an edge gatewaywhere the compressed data is reconstructed and classifiedto identify various abnormal events.

An energy-aware data reduction method form edge-assisted smart healthcare systems adopting AI methodsis proposes in [5]. This system using body area network(BAN), wireless sensor network (WSN), and wearable de-vices to collect the vital signs. The fast error boundedlossy compression pattern (SZ) has been utilized for featureextraction in wearable devices. Afterward, the compresseddata are transferred to a local PC for further processing andanalysis. Researchers have deployed a part of DL on edgedevices and the training network at the cloud. Moreover, aFeed-Forward Neural Network (FFNN) is deployed in anedge device (PC) to classify the compressed data, whichhelps with detecting the stress level for the case study.Researchers in [10] have proposed a real-time smart re-mote monitoring framework using edge technology and AImethods. The proposed framework consists of four layers,including the sensing layer for data collection, edge devicelayer for offline pre-processing, edge server layer, and cloudlayer for further online processes. The offline classifica-tion methods are trained in the cloud for predicting thehealth state of the patient in disconnected emergency events.The proposed method adopts whale optimization algorithm(WOA) and naıve Bayes classifier (NB) for choosing theslight set of features with high accuracy level.

3.2.2 AI in edge servers

In [4], the authors have proposed an edge-based smarthealthcare system to detect severity and transmit alerts forcardiovascular diseases. Leveraging ML-based algorithmslike rapid active summarization for effective prognosis(RASPRO) and support vector machine (SVM), the pro-posed system has conducted data compression, data fusion,and data classification in an android smartphone as an edgedevice. Moreover, the authors in [2] have introduced a newmodel for controlling the outbreak of the Chickungunyavirus. The proposed smart healthcare system leverages edgetechnology as a bridge between cloud storage and end-users. BANs and smartphones are adopted to collect dataand transfer them to the edge servers where ML-basedfuzzy k-nearest neighbor (FKNN) algorithm classifies thedata coming from edge devices. The results are sent tocloud data centers. Also, in [3], the authors have proposeda new smart healthcare framework that utilizes edge tech-nology and DL technique for automatically analyzing andpredicting cardiac diseases. This system delivers a smarthealthcare service as a fog service. The proposed model in-cludes body area networks (sensors), edge gateways (smart-phone), and Fog-Bus module that consists of broker node

5

(Laptop), worker node (Raspberry Pi), and cloud data cen-ter (Microsoft Azure). The set partitioning in hierarchicaltrees (SPIHT) and singular value decomposition (SVD) tech-niques are adopted for data filtering in the broker node, andan ensemble DL technique is used for data classification inthe worker node.

Researchers in [7] have designed an automatic epilep-tic seizure prediction and localization system exploitingedge technology and AI techniques. They have proposed anovel multimodal data processing method on mobile edgecomputing environment using convolutional deep learning(CNN) and wavelet transform for extracting high order fea-tures and a nonlinear SVM classifier along with generalizedradial basis function (GRBF) kernel and long short-termmemory (LSTM) DL for classifying the extracted electroen-cephalogram (EEG) and DL features.

Also, in [9], a real-time data analysis solution for medicaldata stream analysis has been performed, which lever-ages edge technology and AI methods for detecting dataanomaly. The proposed architecture comprises four differentlayers. The sensor layer fetches and dispatches medical datausing wearable devices. The pre-process layer adopts rasp-berry pi to convert the collected raw medical data streamsinto resource description framework (RDF) streams. In thecluster processing layer, a hierarchical temporal memory(HTM) algorithm is employed to cope with the anomalydetection issue. The last layer, persistent layer stores the an-alyzed data to provide the possibility of further evaluation.

In [8], authors have designed an edge-based data dis-semination model in which the mobile edge servers areutilized for disseminating various data segments. The pro-posed model includes the installation method and datapropagation technique that are respectively responsible forensuring the integrity of data objects and preventing thetransmission collision. Moreover, in order to provide ef-ficient data dissemination method in heterogeneous edgenetworks, researchers have used adaptive protocol and thek-means clustering algorithm in this research.

Researchers in [15] have suggested a task offloading andresource allocation model for edge-assisted smart healthcaresystems using AI methods. The authors have considerededge computing with three layers, including devices layer,edges layer, and clouds layer. Cloudlets with micro datacenters are deployed at the edge of the network as edgeservers. The proposed model uses a linear regression algo-rithm (LRA) at the edge servers to analyze the user requestsand a pipeline tree classifier (PTC) for clustering the jobs.Moreover, fruit fly optimization algorithm (FOA) is used tooptimize the resource allocation procedure.

3.3 Discussion and Summary

With regard to the investigated papers, most of the re-searchers have adopted AI on edge technology to managethe collected medical data. In most cases, data managementcontributes to reducing the amount of transmitted data.Since a lower amount of data needs to be conveyed to thecloud, cost, response time, latency, and network congestionwill be decreased, bandwidth and energy usage will bediminished, and accuracy and reliability will be enhanced.Also, the smart healthcare system will be able to makedecisions swiftly. The last column of Table 1 provides more

information about AI benefits in edge assisted smart health-care solutions. Th ensuing section discusses a new systemmodel which tries to provide a single system covering mostof the mentioned AI contributions and benefits.

4 SYSTEM MODEL

This section describes a novel smart healthcare frame-work leveraging edge technology and employing differentAI techniques like DRL. Figure 4 demonstrates the proposedsystem model with three main layers, including sensor layer,edge layer, and cloud layer. In this model, the end-user canbe a patient at a smart home being monitored remotely, apatient in a health institute, elderly people, or even a healthyperson who cares for his fitness. Caregivers can be any of ahealth professional, a health institute, a mobile ambulance,family members, or a drugstore.

The proposed system adopts different AI techniques,including DRL, distributed from sensors to the cloud, tofacilitate smart healthcare services. Being flexible, the pro-posed framework can automatically define a proper solutionmap for the encountered issue. Based on the size of thefaced problem and the volume of possible collected data,it can distribute the processes from sensors to cloud or fromedge to cloud. Regarding this, various layers of DRL can bedeployed on different layers of the proposed framework. Asa case in point, should the purpose of the system is fitnesstracking for a healthy person, there is no need to performpart of processes on the cloud. Therefore, the system canaccomplish processes on sensors and edge layer. The resultscan either be temporarily stored on edge or permanently onthe cloud according to the end-user’s request. The system isflexible enough to choose the appropriate place for hostingthe input and output layers of DRL, depending on thecurrent healthcare issue.

The sensor layer is typically responsible for collectingdata from the end-user and transferring the raw recordsto the edge layer, either to the edge devices or directlyto the edge servers. We believe that compressing data inthe sensor layer helps the system to reduce the overalllatency, network congestion, and computation complexityon the edge layer. The data compression can be handledby adopting lightweight AI techniques such as CS on thesensors.

The edge layer consists of edge devices and edge servershosting AI techniques. Many research attempts have beendone in the field of smart healthcare leveraging AI andedge technology. However, there are still some unaddressedchallenges. For example, DL needs a vast volume of datato train the algorithms. But, in some cases such as Tele-surgery, there is not enough data available to be trained.Also, the trained data for a specific disease using DL is notadoptable for another similar illness; thus, the data shouldbe re-trained. Moreover, RL, as an autonomous self-trainingsystem, should make several small decisions to achieve asophisticated objective. In case of complex decisions withmany states, like monitoring a mobile patient with a chronicdisease which may have countless states due to the unstablecontext information of a moveable patient, the RL algorithmmay not be able to handle the problem since learning fromall states and regulate the reward path can be uncontrollablefor the algorithm. A potential solution for these limitations

6

TABLE 1: Summary of AI-based edge assisted smart healthcare solutionsR

efer

ence

Algorithm AI Category AI contribution AI deploy-ment place

Healthcareapplicationcase

Main objectiveof the system AI benefits

[6] RELM DL Data classifica-tion

Edge device(Raspberry Pi)

Activity moni-toring

Human activitydetection

Low communication latency,swift decision making, and lowcost

[1] DL(SAE) +ML Hybrid

Data compres-sion (featureextraction), dataclassification

Edge device(smartphone)

Neurologicaldisorders

Epilepticseizuredetection

Low bandwidth usage, low en-ergy usage, low response time,and high reliability

[11] DL (RNN) DL Predicting the ac-tivities

Edge device(laptop)

Activity moni-toring

Human activityprediction

Real-time supporting and highprediction rate (99.69%)

[12] DL (DNN) DL Stress analysis Edge device(Smartphone)

Behaviorrecognition

Stress Level De-tection

High detection accuracy(99.7%), rapid detection, andreal-time supporting

[13] RL RL Data privacy Edge device Activity moni-toring Data privacy

High privacy, low energy con-sumption, and low computa-tional latency

[14] CS,ENN/KNN Hybrid

Data com-pression, dataclassification

Sensors, Edgegateway(ODROIDXU4)

Cardiovasculardisorders

Resourceallocation

Low execution time, high accu-racy, low latency, and real-timesupporting

[5] SZ algorithm,FFNN Hybrid

Data com-pression andclassification

Edge devices(PC, wearabledevice)

Behaviorrecognition Data reduction High lifetime, low energy con-

sumption, and high accuracy

[10] NB, WOA HybridFeature selectionand data classifi-cation

Edge devices Cardiovasculardisorders

Remote moni-toring

Real-time supporting, big dataanalysis, and high accuracy

[4] RASPRO,SVM Supervised

Data com-pression, datafusion, and dataclassification

Edge gateway Cardiovasculardisorders

Severitydetection andalerts trans-mission forcardiovasculardiseases

Low bandwidth usage, highavailability, and low cost

[2] FKNN Supervised Data classifica-tion Edge servers Infection out-

break

Control theChickungunyavirus infection

Real-time supporting, low exe-cution time, and low latency

[3] NN DL Data classifica-tion

Worker nodein Fog (rasp-berry pi)

Cardiovasculardisorders

AutomaticHeart Diseaseanalysis

High prediction accuracy andlow execution time

[7] CNN, SVM HybridFeatureextraction, dataclassification

Edge server Neurologicaldisorders Data analysis

High accuracy, high precision,high sensitivity, and reduce thenetwork congestion

[9] HTM Unsupervised Data anomalydetection Edge server Not

mentioned Data analysis High accuracy

[8] K-means Unsupervised Data dissemina-tion Edge server Not

mentionedDatadissemination High reliability, high accuracy

[15] LRA, DTC,FOA Hybrid

Resourceallocation,classificationthe requests

Edge server(cloudlet)

Notmentioned

Resourceallocation

Low energy consumption, lowexecution time, and low cost

can be an AI technique that uses both estimation and map-ping the states together. In fact, the proposed frameworkcan use a real-time self-training model or training fromdatasets depending on the encountered circumstances. Thischaracteristic makes the system flexible to be used for vari-ous issues based on the demands of end-users, like remotemonitoring, self-monitoring, or clinical monitoring. Table 2summarizes the pros and cons of the common AI techniquesutilized in the smart healthcare domain. Hence, consideringthe result of Table 2 and the discussion, DRL is seen as thebest choice since it can cover the cases mentioned above.According to the issue size, the system can automaticallydecide to map all the states/ use estimation or both together.

In case of critical health issues, real-time alert sendingand delivering health services in-time is so crucial to save aperson’s life. Hence, reducing latency, and time of decision

making are vital in smart healthcare systems. The abovementioned flexible AI technique, DRL, can be adopted inparallel on the edge layer by distributing data and processeson various devices or different hidden layers of DRL. It iscrystal clear that the distributed parallel model contributesto more reduction of latency, and more swift decision mak-ing in comparison to an unparalleled model. Since theprocesses are performed faster than the prior methods,the power consumption can also be diminished noticeably.Therefore, the lifetime of battery-powered devices will beincreased.

The medical data, either compressed or raw, are trans-ferred from sensors to edge layer where they are locallystored, classified, and analyzed adopting an AI techniquein parallel mode to deliver real-time decisions for crit-ical healthcare requests. Soon after making a decision,

7

Edge

Cloud

Local data storageEnd-user (patient)

Caregivers

Sensor

AI

AI

Infe

renc

e an

d se

lt-t

rain

ing

part

s of

AI

Infe

renc

e pa

rt o

f A

IT

rain

ing

part

of

AI

Data acquisition01

Data compression02

Applications

Remote data analysis

Long-term data storing

AI Assissted

Real-time local

decision making

Real-time local data analysis

Real-time local data processing

Real-time critical alert

sending

Applications (in Parallel)

Raw dataCompressed data

AI

Benefits

Benefits

Real-time swift Alert sending

Low computational time

Low energy consumption

Low network congestion

Automated healthcare services

Swift response

Increased privacy & security

Low latency

Low network loadLo

w la

tenc

y

Low complexity

Security & privacy

Real-time local data

classification

Fig. 4: An Overview of the System Model

the corresponding alert would instantly be sent to thecaregivers/end-user or both. Thus, they can take actionin time and prevent a possible catastrophe. Exploiting au-tonomous and smart self-training AI methods lied on edgetechnology, the proposed model can offer an automatedsmart healthcare system that can properly work even with-out human intervention.

Furthermore, considering the critical nature of IoT de-vices and medical data which are vulnerable to be at-tacked by hackers, security and privacy management arein the list of vital concerns in smart healthcare systemsand should be taken into account. In our vision, securityand privacy management are considered in different levelsdistributed from sensor to cloud layer adopting automatedAI techniques like DRL. The proposed system model cancover crucial issues like network security, data security, dataprivacy, user privacy, communication privacy, encryption,and authentication of users to prevent illegal access to thesystem.

Eventually, the cloud layer performs remote and furtherdata analysis and long-term data storing. With regards tothe constraint resources of IoT and edge devices, in case oftoo complex processes, the cloud can be the best place forconducting further analysis that is not possible to be doneon edge. For instance, should the smart healthcare system to

train a vast amount of data, the cloud can be the best choicefor this purpose.

In summary, the smart components of the proposedsystem model adopt sensors to collect data related to thereal-time patients’ status while all integrated with physicalitems. All the components are connected to a cloud-orientedsystem which collects and processes all the data producedby sensors. The proposed system model can be consideredas a digital twin system which adopts real world and real-time data to perform simulation leveraging edge intelli-gence and wireless communication, that can estimate how aprocess can be established to get the better results in termsof predicting a disease in advance or saving lives in criticalcircumstances. Moreover, the proposed system model hasa reasonable complexity since there are not any complexalgorithms used in the model. The overall complexity of theproposed system model depends on the health issue andthe amount of data. However, it cannot be greater than thecomplexity of DRL itself.

5 CHALLENGES AND FUTURE RESEARCH DIREC-TIONS

According to our investigation, the ensuing open issuesand challenges in the adoption of AI in edge-assisted smarthealthcare solutions are detected. Future research directionsin this area include both technical and non-technical chal-

8

TABLE 2: Merits and demerits of common AI techniques in smart healthcare domain

AI Technique Benefits Drawbacks

ML Improves over time, Adapt without humanintervention

Needs high amount of time and resources for learning, High level oferror susceptibility

RLHigh learning speed, Real-time response,Dynamic, Capability of learning from smalldata

Lack of clear interpretability, Lack of Integration of Prior Knowledge

DL High dimensional (with many layers), Han-dle large amounts of data

It just provides approximate statistics not accurate results, Needs tonsof data for training, Low speed evolution, Data quality (medical dataare highly ambiguous and heterogeneous; thus, training phase mayface some problems like data sparsity, redundancy, and missing value),Transferring it to the other problems is not easy

DRLAttain complex smart healthcare states,best-accuracy, powerful model, scalability,complex decision-making,

Potential dead-zones, may incur complexity states, off-policy learning,heavy computations, continuous data adjustment

lenges summarized as follows. The first three challenges areconsidered as non-technical and the rest of challenges areknown as the technical challenges.

5.1 Sensor Usage

The most important issue in terms of sensor adoption ishow patients comfort in using sensors. Sensors can beeither implanted or wearable. The elderly people may havedifficulties in using wearable ones. Some others may refuseto adopt the implanted sensors. This challenge also implies apersonal comfort zone, in other words, what seems comfort-able to one patient cannot be assumed the same for another.

5.2 Alert Receiving

The most salient concern in alerting is how family mem-bers/trusted relatives/caregiver are reliable and responsivein taking reaction once they get an alert. For instance,they may not get the alerting notification in time due todifferent reasons (e.g., the dead battery of their device, beingbusy with other tasks, being far from the patient). Theseissues may result in irreparable losses. Thus, fault tolerancemeasures need to be considered.

5.3 Patient’s concerns

Some patients are concerned about the treatment procedureand how they can trust a digital system for distinguishingor predicting their diseases. Thus, they usually resist usingsmart technologies, since they do not have enough informa-tion about how the system works and how doctors interactwith the system and indirectly with the patients.

5.4 Data Loss

Although efficient data compression contributes to reducingthe amount of transmitted data, which may diminish thelatency, power usage, and bandwidth usage, data integrityand information loss as a result of data compression is stilla crucial open issue for smart healthcare solutions.

5.5 Lightweight AI Techniques

Since edge computing apparatuses, which are mostly usedin smart healthcare applications, consist of constraint re-sources and limited power supply, there is a need forlightweight AI techniques that can be executed on them.Although AI techniques usually contribute to obtaininghigh accuracy, in some cases, it is achieved at the expenseof augmented computational cost. An effective resourceutilization may lead to the enhancement of lightweight AI

techniques, which can help the smart healthcare system byanalyzing enormous volumes of medical records producedcontinuously. RL techniques, in comparison to other AImethods, need lower memory resources and computationalcomplexity. The RL techniques usually learn from their priorexperiences; then, they do not need training with datasets.Thus, using RL can lead to preserving computational re-sources and time as well, which are considered as the criticalcriteria in the healthcare systems.

5.6 Redundant Training Data

The DL techniques adopt a huge sample space of datasetsfor training. Each sample consists of various features thatare adopted in the detection/ prediction of diseases. Never-theless, there is substantial redundancy in samples transfer-ring almost identical information. Hence, the computationalresources should work on processing similar informationmany times, which increases the time and computationalcomplexity. These two factors are so essential in case ofreal-time critical medical situations to save the patient’slife. Therefore, proposing a strong and efficient algorithmfor diminishing redundant training and providing effectiveresource utilization still is a challenging issue.

5.7 Privacy and Security

Since data is stored close to the end-users thanks to the edgetechnology, individual users own the data entirely and canprotect them. However, due to the constraint resources ofusers’ devices, they cannot benefit from strong privacy andsecurity protection algorithms. Thus, there is still consider-able room for enhancing the privacy and security of smarthealthcare systems leveraging edge technology and AI. Alightweight AI technique for coping with this issue can be apromising research area.

5.8 Automatic Network Management

With regard to the critical nature of smart healthcare sys-tems, it is so vital for these systems to have the capabil-ity of real-time and automatically managing the networkissues. The adoption of lightweight AI techniques such aslightweight RL can lead to smart and robust network trafficmanagement in smart healthcare systems. Also, in orderto prevent the security gaps and learn the network forautonomously diminishing any possible attacks, the smarthealthcare system can utilize an ML-based intrusion de-tection system. Moreover, software-defined network (SDN)has a crucial role in the implementation of the autonomous

9

network management. The use of learning algorithms canhelp to control the virtualization of network componentsand functions to deliver a self-regulated smart healthcaresystem.

6 CONCLUSIONThis article illuminates how edge technology, along with

AI techniques, can enhance the smart healthcare systems.Processing locally, edge technology improves smart health-care systems by reducing latency, network burden, andpower consumption. Also, AI injects the smartness to thesystem. The integration of AI and edge technology makesthe components of smart healthcare systems smarter andbrings many advantages to them. We introduce a novelsmart healthcare framework that leverages different AI tech-niques in a parallel mode on edge technology. The proposedmodel distributes the processes from sensors to the cloudservers. Conducting some processes using a lightweight AItechnique on sensors results in diminishing the latency, com-plexity, and network load. Furthermore, adopting parallelprocessing on edge layer employing a flexible AI methodcontributes to more reduction of latency, swift responseand decision making, and sending real-time alerts to thecaregivers, which are so vital in smart healthcare services.The proposed model considers security and privacy require-ments for the entire smart healthcare system, including thesensitive medical data, the associated network connections,and users’ access to the smart healthcare system. Althoughthe proposed system model covers some limitations of priorsmart healthcare frameworks, there are still other unad-dressed challenges that need to be addressed, comprisingdata loss, and autonomous network management.

REFERENCES[1] A. A. Abdellatif, A. Mohamed, C. F. Chiasserini, M. Tlili, and

A. Erbad, “Edge computing for smart health: Context-aware ap-proaches, opportunities, and challenges,” IEEE Network, vol. 33,no. 3, pp. 196–203, 2019.

[2] S. Rani, S. H. Ahmed, and S. C. Shah, “Smart health: a novelparadigm to control the chickungunya virus,” IEEE Internet ofThings Journal, vol. 6, no. 2, pp. 1306–1311, 2018.

[3] S. Tuli, N. Basumatary, S. S. Gill, M. Kahani, R. C. Arya, G. S.Wander, and R. Buyya, “Healthfog: An ensemble deep learningbased smart healthcare system for automatic diagnosis of heartdiseases in integrated iot and fog computing environments,”Future Generation Computer Systems, vol. 104, pp. 187–200, 2020.

[4] R. K. Pathinarupothi, P. Durga, and E. S. Rangan, “Iot-based smartedge for global health: Remote monitoring with severity detectionand alerts transmission,” IEEE Internet of Things Journal, vol. 6,no. 2, pp. 2449–2462, 2018.

[5] J. Azar, A. Makhoul, M. Barhamgi, and R. Couturier, “An energyefficient iot data compression approach for edge machine learn-ing,” Future Generation Computer Systems, vol. 96, 2019.

[6] M. Al-Rakhami, A. Gumaei, M. Alsahli, M. M. Hassan, A. Alamri,A. Guerrieri, and G. Fortino, “A lightweight and cost effectiveedge intelligence architecture based on containerization technol-ogy,” World Wide Web, pp. 1–20, 2019.

[7] M.-P. Hosseini, T. X. Tran, D. Pompili, K. Elisevich, andH. Soltanian-Zadeh, “Multimodal data analysis of epileptic eegand rs-fmri via deep learning and edge computing,” ArtificialIntelligence in Medicine, vol. 104, p. 101813, 2020.

[8] C. Shu, Z. Zhao, G. Min, and S. Chen, “Mobile edge aided datadissemination for wireless healthcare systems,” IEEE Transactionson Computational Social Systems, vol. 6, no. 5, pp. 898–906, 2019.

[9] L. Greco, P. Ritrovato, and F. Xhafa, “An edge-stream computinginfrastructure for real-time analysis of wearable sensors data,”Future Generation Computer Systems, vol. 93, pp. 515–528, 2019.

[10] M. K. Hassan, A. I. El Desouky, S. M. Elghamrawy, and A. M.Sarhan, “A hybrid real-time remote monitoring framework withnb-woa algorithm for patients with chronic diseases,” Future Gen-eration Computer Systems, vol. 93, pp. 77–95, 2019.

[11] M. Z. Uddin, “A wearable sensor-based activity prediction systemto facilitate edge computing in smart healthcare system,” Journalof Parallel and Distributed Computing, vol. 123, pp. 46–53, 2019.

[12] L. Rachakonda, S. P. Mohanty, E. Kougianos, and P. Sundaravadi-vel, “Stress-lysis: A dnn-integrated edge device for stress leveldetection in the iomt,” IEEE Transactions on Consumer Electronics,vol. 65, no. 4, pp. 474–483, 2019.

[13] M. Min, X. Wan, L. Xiao, Y. Chen, M. Xia, D. Wu, and H. Dai,“Learning-based privacy-aware offloading for healthcare iot withenergy harvesting,” IEEE Internet of Things Journal, vol. 6, no. 3,pp. 4307–4316, 2018.

[14] H. Djelouat, M. Al Disi, I. Boukhenoufa, A. Amira, F. Bensaali,C. Kotronis, E. Politi, M. Nikolaidou, and G. Dimitrakopoulos,“Real-time ecg monitoring using compressive sensing on a hetero-geneous multicore edge-device,” Microprocessors and Microsystems,vol. 72, p. 102839, 2020.

[15] K. Lin, S. Pankaj, and D. Wang, “Task offloading and resource allo-cation for edge-of-things computing on smart healthcare systems,”Computers & Electrical Engineering, vol. 72, pp. 348–360, 2018.